Codebase Audit And Roadmap

Codebase Audit & Roadmap

When a user names a project and asks "what's in it" / "what's next" / "make a PRD" / "audit this" — this skill covers the full discovery-to-plan workflow.

When to Use

• User says "check this repo" or "audit this project"
• User names a project and asks for a feature summary or to "make a PRD"
• User asks for a gap analysis before a launch/rebuild
• User references a project you haven't encountered before and needs structured analysis

Workflow

Phase 1 — Locate the Project

1. Ask / clarify what they mean — get a name, repo URL, or GitHub handle
2. Search locally: find ~ -maxdepth 4 -name ".git" -type d
3. Search GitHub: use web_search with site:github.com <handle> patterns. Try known handles (S3YED, weblyfe, etc.)
4. Try multi-case variations — the exact case matters on GitHub. If user says "wai", search S3YED/wai, s3yed/WAI, etc.
5. Clone private repos via SSH if SSH key is configured (ssh -T git@github.com to verify)
6. Fallback: ask for the exact URL if web search fails

Phase 2 — Codebase Analysis (Systematic)

Read in this order, top to bottom:

1. README.md — overall purpose, features, stack, setup
2. AGENTS.md / CLAUDE.md / SKILL.md — project-specific instructions for agents
3. package.json — dependencies, build scripts, test commands (run npm ls 2>/dev/null | head)
4. Git log — git log --oneline -10 --all — recent activity, branch structure
5. Project structure — find src -maxdepth 2 -type d | sort, or list top-level files
6. Key feature files — read every component/API/hook file systematically:
   • api/*.ts — serverless endpoints (feature surface)
   • components/*.tsx — UI screens (feature surface)
   • utils/*.ts — shared logic
   • hooks/*.ts — custom hooks
   • migrations/*.sql — DB schema
7. Types/interfaces — the data model tells you the domain
8. Tests — npm test or vitest run — test coverage is a health signal
9. Docs/ directory — PRDs, architecture docs, integration guides
10. Existing PRDs (docs/*PRD*.md, docs/*PLAN*.md) — read first to compare intent vs reality

Phase 3 — Deployment Audit (SaaS projects)

For Vercel-deployed apps:

1. cd <project> && vercel link --yes to link the local clone to Vercel
2. Read .vercel/project.json for projectId + orgId
3. Check deployments: vercel list --prod or vercel list
4. Check env vars: vercel env ls
5. If CLI is limited, use Vercel REST API:

   curl -s -H "Authorization: Bearer $VERCEL_TOKEN" \
     "https://api.vercel.com/v1/projects/<projectId>/env?teamId=<orgId>"
   

   (Write response to file, parse with python3 to avoid shell quoting issues)

For Supabase projects:

• Look for SUPABASE_URL in env files or Vercel env
• Check if migrations/000_base_schema.sql has been run

Phase 4 — Feature Inventory + Gap Analysis

For every feature found, classify:

Build a table with: Feature | Status | File(s) | Notes

Phase 5 — Produce PRD / Build Plan

Structure the output PRD as a markdown document in docs/:

# [Project Name] — [Launch Name] Build Plan

## 1. Current State Assessment

[Feature table from Phase 4]

## 2. [Phase Name — e.g. P2: Voice Notes]

### What already exists
[Code that exists but may not be wired]

### What to build
[Items with file paths, clear deliverables]

## 3. Build Order

Use a priority system:
- **[Phase 1 — Launch-ready]**: P0 items that must work for go-live
- **[Phase 2 — Production Hardening]**: P1 items for reliability/security
- **[Phase 3 — Scale]**: P2 items for multi-tenant scale
- **[Phase 4 — Polish]**: P3 nice-to-haves

## 4. One-Command Vision

[If applicable: describe the ideal provisioning flow]

## 5. Decisions Needed

[Questions for the user/operator]

Save to docs/<PROJECT>-LAUNCH-PLAN.md and present a concise summary in the reply.

Pitfalls

1. Don't assume features work just because the code exists. A component may be dead code (no import/trigger). An API may exist but the state that feeds it may never be populated. Always trace the data flow: where is the state set, where is it consumed.
2. Demo-mode == broken, not working. If an app silently falls back to fake data when env vars are missing, call it a stability bug, not a feature.
3. Rate limiting on serverless is always in-memory until proven otherwise. Vercel cold starts reset in-memory Maps.
4. Feature-rich does not mean deployed. A project can have massive feature surface but zero deployments on Vercel.
5. Private repos won't show in web search. Try cloning with git@github.com:<handle>/<repo>.git after verifying SSH auth.
6. Shell quoting with env vars containing special chars is fragile. Write Python scripts to files for complex API calls; avoid interpolating env vars directly in shell commands.
7. Verbal feature names may not match repo names. "WAI" = WhatsApp Intelligence, not a tool called "wai". Ask for clarification.===ME:codebase-audit-and-roadmap